What it does

This repo implements a multilabel LSTM classifier that reads time-series patient data—think heart rate, blood pressure, oxygen saturation—and predicts which of several diagnoses apply. It targets the MIMIC-III ICU dataset, the messy gold standard for clinical machine learning. The code is delivered as a Jupyter notebook, not a library: you run cells, inspect outputs, adapt to your own cleaned data.

The interesting bit

The paper itself predates the transformer era, yet the problem remains stubbornly relevant. ICU data is irregular, missing, and multilabel by nature—patients rarely have exactly one thing wrong. The LSTM architecture here is the boring part; the value is in how it handles the temporal structure without requiring perfectly aligned time steps.

Key highlights

• Reimplements “Learning to Diagnose with LSTM Recurrent Neural Networks” (arXiv:1511.03677)
• Built for Python 3.5-era TensorFlow, with minimal dependencies (NumPy, Pandas)
• Designed around the MIMIC-III benchmark pipeline, though the author notes this isn’t the exact dataset the original paper used
• Notebook format makes the training loop and data plumbing visible, not hidden behind abstractions
• Explicitly invites pull requests for alternative datasets or fixes

Caveats

• The README warns this isn’t the original authors’ dataset; reproducibility of paper results is unclear
• Python 3.5 and the TensorFlow version implied are well past end-of-life; expect dependency archaeology
• No training curves, accuracy numbers, or pre-trained weights provided in the README

Verdict

Worth a look if you’re teaching or studying clinical deep learning, or need a concrete starting point for MIMIC-III LSTM baselines. Skip it if you want production-ready code or modern PyTorch implementations—this is a reference implementation frozen in 2016 tooling.